Agents and Skills Overview
In one sentence: Make the model not just "answer" but "act": calling tools, executing multi-step tasks, loading skills on demand; how to train and organize these capabilities is the theme of this section.
Status
🚧 This page is a placeholder outline; the full text has not been written yet.
Section Map
Subtopics
| Page | Question it answers |
|---|---|
| Tool-Use Training | How to teach a model to issue function calls correctly |
| Agent Skills | How to package domain knowledge into reusable skills |
| Agentic RL | How to train multi-turn interactive tasks with RL |
| Multi-Agent | How multiple agents divide work and collaborate |
TODO
- [ ] How agent training relates to the previous sections (SFT/RL): same algorithms, different data and environments
- [ ] Survey of evaluation benchmarks (SWE-bench, τ-bench, etc.)